Development of an efficient fractal based texture analysis technique for improved classification of dermoscopic images

Dermoscopy, a non-invasive imaging technique has been significantly used by the doctors and radiologists for the early diagnosis of the various skin disorders. The characteristically similar nature of the melanocytic skin lesions specifically melanoma and dysplastic nevi make the diagnosis more subjective and time consuming, even for expert clinicians. Computer aided diagnostic system has a great impact on the notable discrimination of two closely similar classes of skin diseases by extracting a large number of effective features. In this reported work fractal geometry has been used for both skin lesion border irregularity measurement and texture features extraction. Here, in fractal based texture analysis technique the dermoscopic images have been decomposed into a set of binary images to extract more effective texture features from different grey regions of the image. From each of the image grey region, some statistical features have been extracted along with the fractal dimension measurement to quantify the intensity variation in that region. In this paper it has been established that the extraction of texture information from different small sub regions of the original image using fractal texture analysis increases the classification performance for both the classes. An analysis of the dependence of the performance of classifier using fractal based texture analysis, on the number of decomposed binary images, has been discussed. The highest classification sensitivity of 93.75% and 91.66% have been achieved for melanoma and dysplastic nevi respectively, using support vector machine (SVM) classifier by extracting fractal based texture features from forty grey scale regions of the dermoscopic images.